Fuzzy-Discernibility Matrix-based Efficient Feature Selection Techniques for Improved Motor-Imagery EEG Signal Classification

Brain activities, called brain rhythms, are the micro-level electrical signals (that is, Electroencephalogram or EEG) generated in our brain while we are performing a task. Even when we imagine a limb movement, it generates the same EEG signals called motor-imagery. Motor-imagery based Brain-computer Interface (BCI) provides a non-muscular means to connect the human brain with limbs through computer-based interpretations. The main aim of this paper is to find a suitable feature-set and a classifier to efficiently classify EEG signals into distinct motor-imagery brain-states. We propose to use sliding temporal window-based approaches for feature extraction from EEG and a mix-bagging classifier which is essentially a bagging-based ensemble of multiple types of learners for motor imagery EEG classification. We observe that mix-bagging with overlapping sliding window-based feature extraction achieves an accuracy of 91.43% on the BCI Competition II Dataset III. To reduce the feature size further, we use a fuzzy discernibility matrix that selects the most discriminative features instead of all the features. This additional feature selection strategy improves the classification accuracy to 92.14% and sets a new state-of-the art result on this dataset.